Building Footprint Extraction for the USA¶

Install package¶

To use the geoai-py package, ensure it is installed in your environment. Uncomment the command below if needed.

In [1]:

# %pip install geoai-py

# %pip install geoai-py

Import libraries¶

In [2]:

import geoai

import geoai

Download sample data¶

In [3]:

raster_url = (
    "https://huggingface.co/datasets/giswqs/geospatial/resolve/main/naip_train.tif"
)
vector_url = "https://huggingface.co/datasets/giswqs/geospatial/resolve/main/naip_train_buildings.geojson"

raster_url = ( "https://huggingface.co/datasets/giswqs/geospatial/resolve/main/naip_train.tif" ) vector_url = "https://huggingface.co/datasets/giswqs/geospatial/resolve/main/naip_train_buildings.geojson"

In [4]:

raster_path = geoai.download_file(raster_url)

raster_path = geoai.download_file(raster_url)

naip_train.tif:   0%|          | 0.00/12.1M [00:00<?, ?B/s]

naip_train.tif:  80%|████████  | 9.71M/12.1M [00:00<00:00, 102MB/s]

naip_train.tif: 100%|██████████| 12.1M/12.1M [00:00<00:00, 103MB/s]

In [5]:

vector_path = geoai.download_file(vector_url)

vector_path = geoai.download_file(vector_url)

naip_train_buildings.geojson:   0%|          | 0.00/456k [00:00<?, ?B/s]

naip_train_buildings.geojson: 100%|██████████| 456k/456k [00:00<00:00, 28.7MB/s]

Initialize building footprint extraction pretrained model¶

The pretained model is adapted from the Esri building footprint extraction model for the USA. Credits to Esri for the model.

In [6]:

extractor = geoai.BuildingFootprintExtractor()

extractor = geoai.BuildingFootprintExtractor()

Model path not specified, downloading from Hugging Face...

Model downloaded to: /home/runner/.cache/huggingface/hub/models--giswqs--geoai/snapshots/92a3d4371b88466e0fc1ab3b0964f45782fca4d0/building_footprints_usa.pth
Model loaded successfully

Extract building footprints¶

Option 1: Extract building footprints as raster¶

In [7]:

mask_path = extractor.save_masks_as_geotiff(
    raster_path=raster_path,
    output_path="building_masks.tif",
    confidence_threshold=0.5,
    mask_threshold=0.5,
)

mask_path = extractor.save_masks_as_geotiff( raster_path=raster_path, output_path="building_masks.tif", confidence_threshold=0.5, mask_threshold=0.5, )

Processing masks with parameters:
- Confidence threshold: 0.5
- Chip size: (512, 512)
- Mask threshold: 0.5
Dataset initialized with 3 rows and 5 columns of chips
Image dimensions: 2503 x 1126 pixels
Chip size: 512 x 512 pixels
CRS: EPSG:26911

Processing raster with 4 batches

  0%|          | 0/4 [00:00<?, ?it/s]

 25%|██▌       | 1/4 [00:11<00:35, 11.71s/it]

 50%|█████     | 2/4 [00:23<00:23, 11.87s/it]

Resizing masks at image edges (set verbose=True for details)

 75%|███████▌  | 3/4 [00:34<00:11, 11.49s/it]

100%|██████████| 4/4 [00:42<00:00, 10.15s/it]

100%|██████████| 4/4 [00:42<00:00, 10.71s/it]

Object masks saved to building_masks.tif

Convert raster to vector

In [8]:

gdf = extractor.masks_to_vector(
    mask_path=mask_path,
    output_path="building_masks.geojson",
    simplify_tolerance=1.0,
)

gdf = extractor.masks_to_vector( mask_path=mask_path, output_path="building_masks.geojson", simplify_tolerance=1.0, )

Converting mask to GeoJSON with parameters:
- Mask threshold: 0.5
- Min object area: 100
- Max object area: None
- Simplify tolerance: 1.0
- NMS IoU threshold: 0.5
- Regularize objects: True
- Angle threshold: 15° from 90°
- Rectangularity threshold: 70.0%
Mask dimensions: (1126, 2503)
Mask value range: 0 to 255
Found 642 potential objects

  0%|          | 0/642 [00:00<?, ?it/s]

 11%|█         | 72/642 [00:00<00:00, 710.06it/s]

 23%|██▎       | 146/642 [00:00<00:00, 724.68it/s]

 34%|███▍      | 219/642 [00:00<00:00, 706.55it/s]

 45%|████▌     | 290/642 [00:00<00:00, 707.76it/s]

 56%|█████▌    | 361/642 [00:00<00:00, 700.28it/s]

 67%|██████▋   | 432/642 [00:00<00:00, 700.33it/s]

 78%|███████▊  | 503/642 [00:00<00:00, 692.96it/s]

 89%|████████▉ | 573/642 [00:00<00:00, 692.38it/s]

100%|██████████| 642/642 [00:00<00:00, 700.45it/s]

Created 623 valid polygons

Object count after NMS filtering: 623
Regularizing 623 objects...
- Angle threshold: 15° from 90°
- Min orthogonality: 30.0% of angles
- Min rectangularity: 70.0% of bounding box area

  0%|          | 0/623 [00:00<?, ?it/s]

 21%|██        | 129/623 [00:00<00:00, 1284.57it/s]

 43%|████▎     | 265/623 [00:00<00:00, 1323.99it/s]

 65%|██████▍   | 404/623 [00:00<00:00, 1352.26it/s]

 89%|████████▉ | 556/623 [00:00<00:00, 1416.45it/s]

100%|██████████| 623/623 [00:00<00:00, 1401.12it/s]

Regularization completed:
- Total objects: 623
- Rectangular objects: 619 (99.4%)
- Other regularized objects: 0 (0.0%)
- Unmodified objects: 4 (0.6%)
Saved 623 objects to building_masks.geojson

Option 2: Extract building footprints as vector¶

In [9]:

output_path = "naip_buildings.geojson"
gdf = extractor.process_raster(
    raster_path,
    output_path="buildings.geojson",
    batch_size=4,
    confidence_threshold=0.5,
    overlap=0.25,
    nms_iou_threshold=0.5,
    min_object_area=100,
    max_object_area=None,
    mask_threshold=0.5,
    simplify_tolerance=1.0,
)

output_path = "naip_buildings.geojson" gdf = extractor.process_raster( raster_path, output_path="buildings.geojson", batch_size=4, confidence_threshold=0.5, overlap=0.25, nms_iou_threshold=0.5, min_object_area=100, max_object_area=None, mask_threshold=0.5, simplify_tolerance=1.0, )

Processing with parameters:
- Confidence threshold: 0.5
- Tile overlap: 0.25
- Chip size: (512, 512)
- NMS IoU threshold: 0.5
- Mask threshold: 0.5
- Min object area: 100
- Max object area: None
- Simplify tolerance: 1.0
- Filter edge objects: True
- Edge buffer size: 20 pixels
Dataset initialized with 3 rows and 5 columns of chips
Image dimensions: 2503 x 1126 pixels
Chip size: 512 x 512 pixels
CRS: EPSG:26911

Processing raster with 4 batches

  0%|          | 0/4 [00:00<?, ?it/s]

 25%|██▌       | 1/4 [00:10<00:32, 10.72s/it]

 50%|█████     | 2/4 [00:22<00:22, 11.11s/it]

 75%|███████▌  | 3/4 [00:33<00:11, 11.17s/it]

100%|██████████| 4/4 [00:41<00:00,  9.94s/it]

100%|██████████| 4/4 [00:41<00:00, 10.35s/it]

Objects before filtering: 684
Objects after filtering: 679
Saved 679 objects to buildings.geojson

Regularize building footprints¶

In [10]:

gdf_regularized = extractor.regularize_buildings(
    gdf=gdf,
    min_area=100,
    angle_threshold=15,
    orthogonality_threshold=0.3,
    rectangularity_threshold=0.7,
)

gdf_regularized = extractor.regularize_buildings( gdf=gdf, min_area=100, angle_threshold=15, orthogonality_threshold=0.3, rectangularity_threshold=0.7, )

Regularizing 679 objects...
- Angle threshold: 15° from 90°
- Min orthogonality: 30.0% of angles
- Min rectangularity: 70.0% of bounding box area

  0%|          | 0/679 [00:00<?, ?it/s]

 21%|██        | 141/679 [00:00<00:00, 1409.65it/s]

 42%|████▏     | 288/679 [00:00<00:00, 1440.76it/s]

 65%|██████▍   | 441/679 [00:00<00:00, 1480.97it/s]

 87%|████████▋ | 590/679 [00:00<00:00, 1483.63it/s]

100%|██████████| 679/679 [00:00<00:00, 1482.94it/s]

Regularization completed:
- Total objects: 679
- Rectangular objects: 648 (95.4%)
- Other regularized objects: 0 (0.0%)
- Unmodified objects: 31 (4.6%)

Visualize building footprints¶

In [11]:

gdf.head()

gdf.head()

Out[11]:

	geometry	confidence	class
624	POLYGON ((455181.6 5277618.6, 455177.4 5277614...	0.994213	1
616	POLYGON ((454990.2 5277628.2, 454990.2 5277627...	0.992599	1
617	POLYGON ((454855.8 5277628.2, 454855.8 5277626...	0.989381	1
278	POLYGON ((454981.8 5277798, 454981.2 5277797.4...	0.986561	1
279	POLYGON ((455052 5277904.8, 455050.8 5277903.6...	0.982763	1

In [12]:

geoai.view_vector_interactive(
    gdf, column="confidence", layer_name="Building", tiles="Satellite"
)

geoai.view_vector_interactive( gdf, column="confidence", layer_name="Building", tiles="Satellite" )

Out[12]:

Make this Notebook Trusted to load map: File -> Trust Notebook

In [13]:

geoai.view_vector_interactive(
    gdf, column="confidence", layer_name="Building", tiles=raster_url
)

geoai.view_vector_interactive( gdf, column="confidence", layer_name="Building", tiles=raster_url )

Out[13]:

Make this Notebook Trusted to load map: File -> Trust Notebook

In [14]:

geoai.view_vector_interactive(
    gdf_regularized, column="confidence", layer_name="Building", tiles=raster_url
)

geoai.view_vector_interactive( gdf_regularized, column="confidence", layer_name="Building", tiles=raster_url )

Out[14]:

Make this Notebook Trusted to load map: File -> Trust Notebook

In [15]:

extractor.visualize_results(raster_path, gdf, output_path="naip_buildings.png")

extractor.visualize_results(raster_path, gdf, output_path="naip_buildings.png")

Using confidence values (range: 0.50 - 0.99)

Visualization saved to naip_buildings.png

Sample visualization saved to naip_buildings_sample.png

In [16]:

extractor.visualize_results(
    raster_path, gdf_regularized, output_path="naip_buildings_regularized.png"
)

extractor.visualize_results( raster_path, gdf_regularized, output_path="naip_buildings_regularized.png" )

Using confidence values (range: 0.50 - 0.99)

Visualization saved to naip_buildings_regularized.png

Sample visualization saved to naip_buildings_regularized_sample.png